The MIDI Archive

Searching the future for what resides in the past ... Explore the sounds of the early web with the MIDI Archive

Archives seem to have come into a kind of fashion as of late. In an age of planetary-scale existential precarity, they seem to offer solace in feeling knowable, certain, and grounding. The training procedures of AI foundation models like GPT have a close relationship with archives too, but instead rely on their accessibility, volume, and givenness, which allows archives to be composed and instrumentalized as training sets.

The MIDI files collected here and used to train the model were once very new. In sifting through them, I've been searching for the feeling of technological transformation in an earlier age, how they combine the possibilties of new aesthetic experiences with the technics of producing and distributing this particular format of media.

I developed this project during my residency at Recurse Center in the fall of 2023 as an exploration of the dynamics between building an archive and building a machine learning model. The media ecosystem that emerges from widespread usage of generative AI feels unknowable at this time, but looking back to the brief window of time when music on the web was dominated by MIDI may tell us something about what drives humans to develop, use, and share technology. In both machine learning and the archive, there exist archetypes of transformation and conservation, which serve as carrier for the hopes and fears about how we ourselves may change.

Technical notes

Before MP3s came to dominate how people would listen to music on the internet, the sounds of the early web (and even BBS and Usenet before the world wide web) were predominantly expressed via MIDI. Its tiny file-size was accomodated by bandwidth limitations of the 1980s and 90s, web-native support for the format came early from browsers like Internet Explorer and Netscape Navigator.

The neural net model used here is simple by design, and you can check out and run the source code yourself right from your browser. My intent is to have the model be able to express something at once whimsical and general about the underlying archive, and to be simple enough to serve educational purposes. It does not represent the state of the art in 2023, nor is it intended to be used seriously as a tool for music creation. Furthermore, I planned for this model to be portable and cheap to run. In its current implementation, it exists as a ~800k parameter decoder-only transformer model that runs from an AWS Lambda Function every day around noon (GMT) to produce a piece of music around three minutes in length. Each time, the model uses tokens generated from the previous pass to produce subsequent tokens, and in this way produces a single continuous piece of music that has no end, but is punctuated by a 24-hour cycle of rest.

This project also owes a debt to BitMIDI, which also developed the MIDI player I've implemented on this site for your listening pleasure.

There are currently 34078 MIDI files in the collection!

In addition to browsing the archive below, click here to listen to music produced by the neural net model. This music is generated daily at noon (GMT), so check back again later!

John Sankey - Harpsichordist to the Internet

This collection of MIDI files is hosted on Dave's J.S. Bach Page, and it covers a portion of John Sankey's MIDI recordings. For a time, he was known as the Harpsichordist to the Internet ...